[ltp] An "inverse" fan problem

[Michael Keyl]

Hi all,

while playing around with the ATLAS library (a highly tuned numerical linear 
algebra library) I have encountered a new annoying problem with the fan of my 
TP R51 (1.7 Ghz Banias cpu): If you torture your cpu really hard (like 
hard-core numerics with all tricks like SSE2 and muladd operations) the fan 
spins not fast enough and did not accelerate early enough to prevent the cpu 
from reaching the "passive" temperature. With a small test programm (taking 
the square of a randomly generated 1000x1000 matrix in an infinite loop) I 
have figured out that the cpu clock goes down from 1.7Ghz to 1.2Ghz (which is 
effectively a performance loss of 25%) after approx. 2 minutes. A more clever 
fan control could (at least) extend this time, because the fan starts to 
accelerate quite late (after the cpu reaches approx. 85C) and it does not go 
beyond 3800 RPM. Is there anybody around with similar problems and/or ideas 
towards a solution?

Michael Keyl